Bhavani Mulakala

Architected and deployed 15+ production-grade Generative AI and LLM applications for payment processing, fraud operations, merchant servicing, dispute management, and regulatory compliance. Built enterprise RAG solutions using LangChain/LangGraph, LlamaIndex, and vector/keyword search systems (Pinecone/FAISS/OpenSearch/Azure AI Search) indexing 25M+ financial documents to reduce hallucinations. Developed multi-agent systems with planning, memory, and tool orchestration (LangGraph, Microsoft Semantic Kernel) to automate payment investigations and reduce manual analyst effort. Fine-tuned Llama 3, GPT-4, Claude, Gemini, and Mistral using LoRA/QLoRA/PEFT and RLHF/SFT to improve QA accuracy and reduce inference costs. Engineered large-scale predictive models for fraud/AML using supervised learning, anomaly detection, and graph analytics, reaching fraud detection accuracy up to 96%. Implemented LLM evaluation and regression validation for groundedness, hallucinations, and retrieval relevance

Designed, developed, and deployed 30+ end-to-end ML models for claims prediction, fraud detection, prior authorization automation, care management, provider network optimization, and member risk stratification. Built predictive analytics platforms handling 75M+ insurance claims annually using supervised/unsupervised/semi-supervised learning, ensembles, deep learning, and anomaly detection. Implemented NLP pipelines for extracting clinical insights from provider notes and medical documents (FHIR/HL7/EMR/EHR) using SpaCy and transformer-based models; built scalable feature engineering on 12TB+ data using Spark/Hadoop/Databricks/Delta Lake. Developed fraud detection and explainability workflows using SHAP/LIME and classical ML methods. Delivered MLOps/AI SDLC with MLflow, Kubeflow, Airflow, Azure ML Pipelines, Docker, Kubernetes/AKS, and CI/CD, reducing deployment cycles from 10 days to under 3 days. Built secure real-time inference APIs supporting millions of predictions per month, and a

Developed end-to-end ML solutions for Nebraska state agencies to improve citizen services, public administration, workforce analytics, healthcare programs, and social service delivery. Built predictive models for Medicaid eligibility and other forecasting/risk use cases using feature engineering, selection, hyperparameter tuning, ensembles, and deep learning techniques. Implemented NLP systems for citizen feedback and document processing (NER, sentiment, topic modeling, clustering) achieving high classification accuracy. Built robust ETL pipelines processing 12TB+ of structured/semi-structured government data using Spark/PySpark/Hadoop/Hive. Automated training and deployment with MLflow, Docker, Kubernetes, Jenkins, and CI/CD; deployed models on AWS services and implemented versioning, monitoring, drift detection, and retraining. Collaborated across stakeholders to deliver production ML solutions while maintaining governance and security controls aligned with state data policies.